minimax-h3宽高比选择

Video models are picky about resolution. MiniMax H3 - the 33B omni-modal video model that landed open-weights in 2026 and does text, image, video, and audio in one context - wants specific width/height combos for a given aspect ratio and pixel budget, and a random "close enough" resolution is a fast way to get a rejection or a distorted generation. MinimaxH3AspectRatio is a tiny node from 🐳 Pond Nodes (comfy_Pond_Nodes) that does one job: turn three dropdown choices into the exact width and height H3 expects.

How it works

This is where it gets interesting, because the node is smarter than a generic calculator. For the most common case - 16:9 at a 32-pixel multiple - it doesn't compute anything at all. It looks up the dimensions in a hardcoded table of reference values taken from MiniMax's own size table: 0.4MP gives you 864×480, 0.5MP gives 960×544, 1.0MP gives 1376×768, all the way up to 2.0MP at 1920×1088. That's the difference between "a number that's roughly right" and "the number the model's docs actually list."

For everything else - other aspect ratios, other multiples, or megapixel values outside the reference table - it falls back to real math: solve for width and height from the aspect ratio and target pixel count, then round up to the nearest multiple of 8, 16, 32, or 64.

The inputs and outputs

Three dropdowns, all self-explanatory:

• aspect_ratio - 9 presets: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:21. Labels carry the 横屏/竖屏 orientation marker so you don't mix up landscape and portrait.
• megapixels - 0.2 through 2.0. The 0.4 default matches the resolution MiniMax's docs steer you toward as the sensible starting point.
• multiple - 8, 16, 32, or 64. H3 wants multiples of 32 for its reference sizes; 8/16 are for the computed fallback cases.

Two INT outputs: width and height. Feed those straight into your H3 workflow's resolution inputs.

Installing it

Same pack install as always - Manager search comfy_Pond_Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Pondowner857/comfy_Pond_Nodes
cd comfy_Pond_Nodes
pip install -r requirements.txt

Restart and it's under 🐳Pond/Tools. It's pure Python math - no models, no GPU, none of the pack's heavy dependencies. (The pack's README does carry a conflict warning for comfyui_HiDream-Sampler; only relevant if your console starts spamming.)

The gotcha

The exact reference-table match only applies at 16:9 with a 32 multiple - that's the case H3's own docs table covers. If you're working in 21:9 or with a 16 multiple, you're getting the computed approximation, which will be close but won't match the published table. Also note this node outputs dimensions only; it has no opinion about frames, duration, or the actual model files. If your H3 workflow still rejects a resolution, check whether it's asking for a specific megapixel tier rather than assuming any number is fine.